Google released Gemini 3.5 Flash yesterday at I/O. The headline that actually matters for builders: a Flash-tier model that beats the previous-generation Pro across coding and agentic benchmarks while running roughly four times faster than other frontier models. The pricing sits at the Flash tier too, which means the cost-vs-frontier calculus that already shifted with Cursor's Composer 2.5 just shifted again.
This is the practitioner's read on what changes, what doesn't, and how I'm rerouting traffic this week.
The Benchmark Picture
The numbers Google published, with the previous-generation Pro for context:
| Benchmark | Gemini 3.1 Pro | Gemini 3.5 Flash | Note |
|---|---|---|---|
| Terminal-Bench 2.1 | mid-60s | 76.2% | Solid, slightly behind GPT-5.5's lead |
| MCP Atlas | n/a | 83.6% | New benchmark, tool-use focused |
| CharXiv Reasoning | n/a | 84.2% | Document/chart reasoning |
| Finance Agent v2 | 43.0% | 57.9% | The single largest jump |
| GDPval-AA (Elo) | 1314 | 1656 | Agentic eval, +342 Elo |
The Finance Agent jump from 43.0% to 57.9% is the one that should make anyone running agentic workflows pay attention. Long-horizon, tool-using tasks are exactly the workloads where the prior-generation Flash was the wrong choice and the prior-generation Pro was the only option. 3.5 Flash collapses that decision down to one model on the basis of capability alone.
Two honest caveats. CursorBench wasn't included in Google's reporting (different harness), so direct head-to-head numbers against Composer 2.5 aren't published. And Terminal-Bench shows 3.5 Flash sitting solidly above the mid-pack but still trailing GPT-5.5 by a few points. Benchmarks are directional, not predictive — validate on your own evals before re-routing real traffic.
The Cost-and-Speed Delta
Pricing is where the strategic decision gets made:
| Model | Input/M | Output/M | Cached input |
|---|---|---|---|
| Gemini 3.5 Flash | $1.50 | $9.00 | $0.15 |
| Gemini 3.5 Flash (non-global) | $1.65 | $9.90 | — |
| Claude Opus 4.7 | ~$15 | ~$75 | — |
| Composer 2.5 (standard) | $0.50 | $2.50 | — |
The headline that Google led with — "within two points of Anthropic's flagship at a third of the price" — undersells the speed delta. 4× the output tokens per second of other frontier models means agentic loops finish in a quarter of the wall-clock time. For interactive coding agents and chat UIs, the latency change is the part users actually feel.
The cached-input price is the lever to watch. At $0.15 per million tokens, repeatedly priming the model with the same long context (a codebase, a long instruction set, a reference document) is essentially free compared to the un-cached path. The tier-routing logic I've written about for Gemini Flash Lite vs Pro needs a refresh — 3.5 Flash now occupies the slot that 3.1 Pro held a month ago, and Flash Lite gets pushed further down the ladder.
When Gemini 3.5 Flash Is the Right Call
Three categories where I'm re-routing this week without waiting for further data:
Long agentic loops
The Finance Agent v2 jump and the +342 Elo on GDPval-AA both point at the same thing: 3.5 Flash holds up over sustained tool-use sessions in a way Flash-tier models historically did not. Production agent loops that previously routed to Pro because the Flash variant was too brittle now route to 3.5 Flash by default.
Cost-sensitive volume
CI jobs, internal tooling, automated review, batch document processing. Anything where I'm running thousands of inferences a day and the per-task cost compounds. The $1.50 / $9.00 pricing, with cached input at $0.15, makes the math obvious. The previous default of "use Flash where good enough, escalate to Pro where it isn't" now starts further down the ladder.
Interactive coding
The 4× speed advantage matters most when there's a human waiting. In-IDE completion, chat-mode coding, agents whose output the user is actively watching. Latency was the place where Flash-tier models always paid for their lower cost. 3.5 Flash no longer pays it.
When the Frontier Still Wins
Three categories where I am not switching:
Top-of-distribution reasoning
For the highest-stakes reasoning tasks — architectural reviews of large codebases, multi-step planning across long horizons, anywhere the cost of a wrong answer outweighs the cost of compute — Claude Opus 4.7 and GPT-5.5 still lead at the very top of the distribution. The 3.1 → 3.5 jump is real, but the frontier is still the frontier on the hardest workloads.
Terminal-heavy work
GPT-5.5 retains the Terminal-Bench lead over 3.5 Flash by a few points. If your agent spends most of its time in a shell — log parsing, debugging, terminal-mediated incident response — that lead is real enough to keep GPT-5.5 primary for those workloads.
Workflows already pinned to a model
Don't migrate a working production agent just because a cheaper, faster model exists. The cost savings disappear the first time a hallucinated migration ships to production. Run shadow traffic, compare on your own evals, and switch only when the gap is small enough to justify the change-management cost — the same migration discipline I've written about for the Gemini family generally.
Migration Planning
A pragmatic rollout for moving production traffic to 3.5 Flash:
- Shadow traffic (week 1). Route five percent of agent traffic through both 3.1 Pro and 3.5 Flash. Log iteration count, tool-call accuracy, latency, and per-task cost. Don't surface 3.5 Flash output to users yet.
- Validate against your own evals (week 2). Benchmarks are directional. Your evals — synthetic traces of tasks that have broken in the past — are what tells you whether the model is ready for your specific distribution.
- Move to standard pricing tier (week 3). Switch the bulk of async and batch workloads. Reserve the faster, more expensive variants only for interactive coding where wall-clock time is user-visible.
- Refresh your routing layer (ongoing). The boundary between "cheap-and-good" and "frontier-only" has shifted. The logic that used to route to 3.1 Pro likely now belongs on 3.5 Flash; the logic that used to route to Flash Lite likely now belongs lower again. Don't leave the routing layer pinned to last quarter's defaults.
What's Not Changed
The unchanging caveats:
- Hallucinations still happen. Less often, but the production validation discipline doesn't change. Schema validation, output review, eval suites — all still load-bearing.
- Benchmarks are not your workload. The numbers Google published are directional. Run your own evals.
- Multi-provider risk is unchanged. A cheaper, faster Google model is not a reason to remove the routing fallback to Claude or OpenAI. Keep the abstraction layer.
- Latency scales with input size. A 1M-context call still does not return in the same wall-clock time as a 5K-token one. UX expectations need to scale with input length.
The Practitioner's Take
The 3.1 → 3.5 jump is the biggest mid-cycle capability shift I've seen from Google in two years. It's also strategically interesting in a way the benchmark numbers don't fully capture: this is Google explicitly betting that the future of model deployment is agentic loops on Flash-tier infrastructure, not chat-style requests on Pro-tier hardware.
That bet looks correct. Most of the agentic work I deploy for clients is bottlenecked on cost per long-horizon run and on the wall-clock time of a multi-step loop. 3.5 Flash improves both. The companies that already adopted the AI-first operating-system framework get a meaningful tailwind from this release without changing a line of their architecture.
The window where 3.5 Flash is the new default and your routing layer is still tuned for last month is small. Refresh it this week.